OCP Global Summit 2025 | Amphenol | Thinking Outside the Rack: Scaling Up with Optical Interconnects

Introduction

At OCP Global Summit 2025, Amphenol presented a new view on optical interconnects in future scale-up architectures.

The talk was given by Amphenol's Director of Standards and Technology, Sam Kosis, and focused on "how to bring more optical links inside the rack and outside the rack, and find the optimal crossover point between copper and optics for short-reach high-speed transmission."

Amphenol's core message was clear:

As AI cluster density and power soar, short-reach optics will no longer be just an extension of high-speed Ethernet but the new backbone of scale-up compute architectures.

Content

1. From PCIe to optical interconnects: the new evolution of scale-up architecture

Amphenol reviewed recent OCP community work on extended PCIe interconnect.

Early study assumptions:

• Copper cables reach 1 m

• Active electrical cables (AEC) reach 3 m

• Optical links extend to 12 m (latency-bound by memory application requirements)

But as retimer technology matured, copper can actually reach 2–7 meters. That means distances once reserved for optics are partly being taken over by electrical links.

However, as rack power density and bandwidth rise further, demand for short-reach optics is re-emerging.

Amphenol noted: "The emergence of short-reach optical interconnects marks our move from the CPU-centric server era into the interconnect-centric accelerated computing era."

2. The architectural logic of scale-up vs. scale-out

Amphenol draws a clear distinction in today's data center architecture:

• Scale-up: integrating compute, memory, storage and GPUs within the same rack into a "composable fabric" via high-speed links.

• Scale-out: connecting many nodes across racks to form large clusters.

The role of optical interconnects is shifting from traditional scale-out (long-reach Ethernet) and gradually penetrating the scale-up layer, becoming the short-to-medium-reach channel connecting GPUs, CXL, PCIe and UALink.

What this means: optical paths in future AI clusters will no longer run only between switches, but will extend into the compute modules and memory architecture themselves.

3. The CPO vs. pluggable divide

Amphenol pointed out that over the past 18 months, the industry discussion has almost turned into a binary debate of "CPO vs. pluggable modules":

• The CPO camp argues:

  • Higher packaging density, lower power, better efficiency.

• The pluggable camp emphasizes:

  • Lower cost, better serviceability, mature volume manufacturing.

Amphenol's conclusion:

"CPO will not fully replace pluggables; the two will coexist depending on SerDes reach (the distance capability of the electrical channel)."

In other words, as long as SerDes can support 1–7 meter electrical links, the point where optics takes over will be pushed out. The real breakthrough must come from defining a new short-reach optical electrical-interface standard.

4. Ethernet vs. PCIe: a precursor to protocol convergence

Kosis showed a chart comparing the evolution of Ethernet and PCIe:

• Ethernet is heading to 200G/lane, but copper reach has shrunk dramatically.

• PCIe maintains longer chip-to-chip links but must rely on retimers.

He proposed an interesting direction:

"We may see an Ethernet + PCIe hybrid fabric that combines Ethernet's modularity with PCIe's flow control and low latency."

This echoes the two major industry trends of UALink and the Ultra Ethernet Consortium (UEC), both of which aim to merge communication protocols with memory semantics (load/store).

5. The key to channel design: from PCB to cable-on-board

As channel insertion loss increases, Amphenol stressed that channel design is undergoing a fundamental shift:

• Traditional PCB traces have hit their limit (close to 30 dB at 200G/lane).

• New designs introduce Copper Cable on Board (CCoB) and hybrid package integration, embedding copper cabling in the board to reduce reflections and insertion loss.

Kosis showed two integration paths:

1. Short-reach hybrid optical-electrical packaging (copper + optics on package): optical and electrical channels within the same footprint.

2. Full optical module conversion: moving the E/O conversion point forward to shorten electrical distance and cut power.

These concepts are the prototype of future NPO (near-packaged optics).

6. Power and rack challenges: half-megawatt racks

Using OCP Open Accelerator 2.0 as an example, Kosis noted:

• Current 8-GPU racks are approaching 50 kW.

• Future high-density racks (about 500 GPUs) will reach 0.5 MW.

This pushes existing copper cable and PCB designs to their limits on cooling and EMI.

Therefore, the coming thermal and electrical bottlenecks become the best opening for optical interconnects to penetrate.

Amphenol's recommendation: in such high-power environments, the electrical channel budget should be redefined, with short-reach optical channels planned proactively to relieve thermal and signal-integrity pressure.

7. A call to action: establish a short-reach optical interconnect standard

Kosis closed with a clear call to action:

"The industry needs a new physical-layer spec for short-reach optics — clearly defining E/O conversion distance and electrical channel length — to truly trigger the tipping point where optics replaces copper."

He suggested that OCP, OIF, PCI-SIG and IEEE jointly define a unified specification supporting short-reach optical-electrical conversion, so that CPO, NPO, pluggables and onboard optics can coexist under the same design language.

Conclusion

Amphenol's view at OCP 2025 is quite pragmatic:

For optics to replace copper, it won't be by competing on cost or performance, but by redefining channels and rebalancing system design.

As rack power, GPU density and bandwidth keep soaring, short-reach optical interconnects will shift from an optional solution to a structural requirement.

The real meaning of "Thinking Outside the Rack" is to break out of traditional rack logic and make optical interconnects part of system design, rather than a bolt-on component.

Extended perspectives

1. Technology impact

   • Amphenol's proposal marks the point where "NPO and short-reach optical interconnects are formally seen as core to scale-up architecture", in sharp contrast with Broadcom's CPO and Credo's copper-cable strategy.

   • If a new short-reach standard emerges, it will create a common platform for multiple light-source technologies such as SiPh, VCSEL and EML.

2. Supply chain observations

   • Amphenol is both a connector and cable giant and a standards driver, giving it a leading voice in future hybrid optical-electrical design.

   • If OCP and OIF jointly set a short-reach optical channel standard, Amphenol could become the bridge company for rack-level optical-electrical connectivity specifications.

3. Market trends

   • "Short-reach optics" will become the next growth focus for the optical communications industry, especially in AI training racks, CXL memory fabrics and GPU backplanes.

   • Between 2026 and 2028, optical modules will move from switches into GPU modules — the biggest paradigm shift in optical applications yet.